Full-Reference and No-reference Image Blur Assessment Based on Edge Information

— Blur images are often subjected to the loss of high frequency content during acquisition, compression and multimedia transmission. Hence, objective blur assessment is implemented to identify and quantify image quality degradation by blurriness artifact in order to maintain and control the quality of the images. In this paper, objective full-reference and no-reference blur assessments using edge information are presented with the aim to provide computational models that can automatically measure the amount of blurriness artifact such as Gaussian blur on the images. The amount of Gaussian blur on an image, also known as the final blur measurement is determined by averaging the sum of edge width over all detected edges which satisfy the edge criteria. The final blur measurement for all test images based on full-reference and no-reference implementations are also validated with subjective results. The validation results show that the objective full-reference and no-reference blur assessments correlate closely to perceptual image quality.


I. INTRODUCTION
Along with the rapid advances in digital and multimedia imaging industry, digital images have been playing an increasingly important role in the communication of visual information. Unfortunately, these images can be degraded during acquisition, compression, transmission or even processing. Thus, a measure to the image quality is necessary to assess the degree of degradation.
The method of measuring the degree of degradation is categorized into objective and subjective evaluations. Objective method is based on mathematical measure while subjective method relies on the perception of a selected group of human observers such as professionals or lay viewers. Objective image assessment is least preferable as the most reliable means to access the image quality is through subjective evaluation, where the quality of the image is evaluated by human. However, it is not an easy task to conduct a subjective evaluation image assessment because it is expensive and time-consuming [1]. Due to the limitations in subjective evaluation, researchers believed that it is useful to design objective method as long as it produces results that correlate closely with human visual system (HVS) [2]. There are 3 approaches to objective image quality assessment that can be considered such as noreference (NR) measurement, reduced-reference (RR) measurement and full-reference (FR) measurement [2]. A FR measurement requires complete availability of the original image while NR method accesses the distorted image without using any reference image. Meanwhile, RR method needs partial information of the reference image and it is a trade-off between both NR and FR methods.
Researchers have developed various perceptual image quality metrics where these metrics are used to measure the global distortion but a perceptual objective image quality analysis to measure a specific artifact such as blur is rarely found. Therefore, this paper aims to propose an objective blur assessment using edge information which correlates well with human visual perception based on FR and NR implementations.
II. BLUR ASSESSMENT Most digital imaging capture devices and electronically displaying visual information devices aim to produce the best image quality. When an artifact like blur is introduced in an image due to acquisition or compression, the blur image has to be enhanced in order to look visually appealing. However, there is a possibility that this corrected image might not satisfy human perception. Various people evaluate the quality differently due to the sensitivity of human eye. Therefore, the blur metric which correlates with human visual as according to P. Marziliano et al. [3] is introduced to measure the blur image to determine the level of degradation so that it can be corrected to a certain extent to produce a better quality image and maintain the pleasure of human observers in viewing the image.
There also other researches in blur assessment such as a No-reference Quality Metric for Measuring Image Blur presented by E. Ong et al. [4] in 2003 using the gradients' directions of each pixel in the image, The Blur Effect: Perception and Estimation with a New No-Reference Perceptual Blur Metric proposed by F. Crete et al. [5] in 2007 by comparing the variations between neighboring pixels while X. Wang et al. [6] published a Blind Image Quality Assessment for Measuring Image Blur in 2008 which is also based on edge gradient information. Although in [4] and [5], the researchers mentioned that their metrics correlate well with results obtained from subjective experiments but the correlation values are not reported. Moreover, in [6], the authors did not mention about validating their metric with subjective scores. However, the blur assessment using edge information presented in this paper which is based on the method implemented by P. Marziliano et al. [3] is highly correlated with subjective blur ratings when measuring on Gaussian blurred images. Based on Pearson linear correlation and Spearman's rank order reported in [3], the metric has shown 96% correlation with its subjective evaluation.
Nevertheless, the blur measurement is applicable to numerous applications such as blur estimation in digital photography, image processing, printing or as a simple metric in comparing two images.

A. Test Images
There are two sets of images used as test images. The first set consists of synthetically generated blur artifact on 29 RGB original images from LIVE database [7] using Gaussian blur lowpass filter of kernel size 15 × 15. Each of the images is blurred using Gaussian standard deviations, σ ∈ {0.4, 0.8, 1.2, 1.6, 2.0}. This set of images is used for blur assessment implementation in Section IV-A until Section IV-C.
In the second set, these test images are used for analysis in Section IV-D. There are 29 original images blurred using circular-symmetric 2-D Gaussian kernel of standard deviations ranging from 0.42 to 15 pixels by H. Sheikh et al. [8] to generate 145 blur images. Each source images has five blur images distorted with different standard deviation.
The total test image in each set is 174 images including original and distorted images. Fig. 1 shows some of the test images used for the blur assessment.

B. NR Implementation
The blur image is converted to YCbCr color space as blurriness is measured on luminance (Y) component. The vertical Sobel edge detector is applied to the Y component image to determine the vertical edges. Blurriness can be measured on gray-level component by converting the RGB test image to grayscale image. The processing steps are illustrated in Fig. 2 below.

C. FR Implementation
The original image is required for this implementation. Both original and blur images have to perform YCbCr color conversion. The vertical Sobel edge detector is then applied on the Y component image from the original image to obtain the location of the edges. As shown in Fig. 3, the blur measurement is based on the Y component of the blur image and edge locations detected from the original image.  Thus, the edge criteria are introduced to overcome this problem. The edge is only accepted based on certain conditions as follows: • An edge must reside in between a local minimum and a local maximum.
An example is shown in Fig. 6 below on calculating the edge width for an edge which falls on position P1. The edge width for this edge is given by the difference between P2 (start position) and P2′ (end position), which is 7 pixels for this example. 2) using grayscale images: This analysis is to compare with the blur measurement in Section IV-C-1. Graphically, the results are identical as seen in Fig. 11. 3) horizontal edge information: The NR blur measurement presented by P. Marziliano et al. [3] considers only vertical edges due to performance reasons. However, the blur assessment can be easily extended to horizontal edge detection. Fig. 12 shows that blur metric based on horizontal edge detection is higher than the metric value from vertical edge detection. This applies to most of the test images.

D. Analysis using LIVE Database Gaussian Blur Images
Below are the blur assessment implementations using blur images from LIVE database [7]. The result of each implementation is validated with subjective testing result (DMOS values) from LIVE database [7]. 1) NR Implementation: The scatter plot in Fig. 13 shows a relatively high correlation between the objective and subjective results with R 2 equivalent to 0.81. R 2 value close to 1 indicates excellent linear reliability and correlation.

DMOS against NBM for FR Implementation
where, BM BLUR denotes the raw blur measurement of blur images and BM ORI is the raw blur measurement of original image. The graph of SDBM against DMOS values is then plotted for every blur level in the image and both lines are aligned using Microsoft Excel trendline function. A gain, α is obtained from the average of DMOS and SDBM line gradients. Thus, the gain is multiplied with SDBM values to acquire the normalized blur measurement (NBM) values.

E. Discussion
Based from the results, the blur metric increases with Gaussian blur standard deviation. The increase of degree of blurriness in an image produces a smoother image, hence this affects the number of detected edges to decrease and the edge width of each detected edge to increase. Thus, from equation (1), the blur metric eventually increases due to the increase of sum of edge width and the decrease of total number of edges.
Moreover, the metric values based on FR implementation are lower than that of the metric values from NR implementation. This is because in FR method, the increase in blur level does not affect the number of detected edges.
In both Section IV-C-1 and Section IV-C-2, the results indicate that the comparison graphs are similar. Meanwhile, Section IV-C-3, it is observed that the horizontal edge detection does not improve the measurements. For the analysis in Section III-D, the results from NR and FR blur assessments correlate closely to subjective results.
Concisely, the blur assessment has low computational complexity. The total computational time for measuring blur on all 174 test images in any method of implementation (NR or FR) is approximately 160 seconds. Hence, the elapsed time of measuring blur artifacts on an image is less than 1 second.
V. CONCLUSIONS A low computational complexity objective blur assessment is successfully implemented to measure the degree of blurriness in an image. Different types of analysis are also carried out to evaluate the performance of the blur assessment. The results of the assessments are as well validated with subjective results where the validation indicates that there is a close correlation between the objective blur assessments with human perception.
As for further work, the blur assessment can be easily extended to digital video quality assessment to measure the level of blur in every consecutive frame in the video. The global blur measurement of a video will be the average of local blur measurement in every successive frame. Hence, the blur assessment can be improved so that is also applicable to real-time application. Besides that, blur identification algorithm is recommended to implement together with the blur assessment. Instead of measuring the amount of blur artifacts that present in an image, the type of blur that degrades the image can be identified so that user will have knowledge on which type of blur distorts the image quality. Finally, further research can be done on the blur assessment by measuring blurriness in color images on other color component such as hue instead of measuring on the luminance component.